Research map / Reinforcement learning / RL from feedback
Humanline: Online Alignment as Perceptual Loss
Sijia Liu, Niklas Muennighoff, Kawin Ethayarajh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use humanline variant on offline data | DPO/KTO/GRPO (Llama3-8B-Instruct) · AlpacaEval2 winrate | winrate: up to 1.6x higher than offline | Improved |
| use humanline variant on online data | DPO/KTO/GRPO · AlpacaEval2 winrate | winrate | No clear effect |
| add humanline syncing (ablation) | offline+humanline variant · AlpacaEval2 winrate | winrate | Improved |
| add humanline clipping on top of syncing (ablation) | offline+humanline variant · AlpacaEval2 winrate | winrate | Improved |
| humanline GRPO with data sampled 64x less often | GRPO (Qwen2.5-1.5B-Instruct) · MATH500 | Pass@1: 0.593 vs 0.593 | No clear effect |
| sample data 64x less often without humanline (ablation) | GRPO (Qwen2.5-1.5B-Instruct) · MATH500 | reward / accuracy | Hurt |
| sync reference every step (k=1) in math (ablation) | humanline GRPO · MATH500 | training reward | Hurt |
| sync every k in [12, 24] in math (ablation) | humanline GRPO · MATH500 | training reward | Improved |
| train on lowest quartile of offline data by token log-prob (ablation) | offline+humanline DPO · AlpacaEval2 winrate | winrate | Hurt |
| use humanline sampling instead of humanline clipping (ablation) | humanline variants · AlpacaEval2 winrate | winrate | No clear effect |
Closest papers by meaning
- Model Alignment as Prospect Theoretic Optimization (ICML 2024 · similarity 0.90)
- Direct Alignment with Heterogeneous Preferences (NeurIPS 2025 · similarity 0.88)
- Synthesizing Images on Perceptual Boundaries of ANNs for Uncovering and Manipulating Human Perceptual Variability (ICML 2025 · similarity 0.87)
- Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences? (NeurIPS 2025 · similarity 0.86)
- Learning to Decide with AI Assistance under Human-Alignment (NeurIPS 2026 · similarity 0.86)
- Understanding the Learning Dynamics of Alignment with Human Feedback (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
